#0: Remove incorrect assertion on cb page sizes being a multiple of 4…

[tt-aho]

… (should have been 16) and update circular buffer init/usage to support any page size for cbs that aren't used by compute. Compute CBs still require 16B multiple for address/sizes

Ticket

Link to Github Issue

Problem description

We were incorrectly asserting for 4B alignment on main for CB page sizes, when we actually needed 16B. However, we only really need 16B alignment when the CBs are used in compute so we can relax this constraint in some cases.

What's changed

Remove the assertion for 4B alignment for CBs. We will pass and store the original values for address, size, page size when used in a DM kernel, and bit shift these values by 4 when initialized in a compute kernel.

Checklist

• Post commit CI passes https://github.com/tenstorrent/tt-metal/actions/runs/12227021124
• Blackhole Post commit (if applicable)
• Model regression CI testing passes (if applicable) https://github.com/tenstorrent/tt-metal/actions/runs/12238505424
• Device performance regression CI testing passes (if applicable)
• New/Existing tests provide coverage for changes

[tt-aho]

@rtawfik01 @pgkeller I haven't added the watcher validation for the compute cb addr/sizes to this pr yet.

In the non-validated case we will bit-shift the addrs/ptrs during cb init to avoid bit shifting at runtime.
We can't validate during CB initialization since we don't know which cbs are used by compute or not yet, that means when we need to validate then we can't actually bit shift during cb initialization, and have to wait until we access these attributes/generate addresses within the compute apis/llks to validate the address is 16B aligned, and then bit shift.

This will touch quite a few compute/llk files. I can try and do that in this pr, or we can split it and add the validation in a separate pr. Adding the validation separately wouldn't make the behaviour any less safe than what is on main, since we were only checking that the page size was 4B aligned on main so it wasn't guaranteeing any correct behaviour anyways.